fix(cuda_graph): zero out_cache_loc_swa on pad and use int32 (hybrid-SWA accuracy fix) (#24743)

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -182,7 +182,7 @@ class DecodeInputBuffers(ForwardInputBuffers):
             seq_lens = torch.full((max_bs,), seq_len_fill_value, dtype=torch.int32)
             out_cache_loc = torch.zeros((max_num_token,), dtype=cache_loc_dtype)
             out_cache_loc_swa = (
-                torch.zeros((max_num_token,), dtype=torch.int64)
+                torch.zeros((max_num_token,), dtype=torch.int32)
                 if is_hybrid_swa
                 else None
             )
@@ -289,6 +289,12 @@ class DecodeInputBuffers(ForwardInputBuffers):
         if bs != raw_bs:
             self.seq_lens.fill_(seq_len_fill_value)
             self.out_cache_loc.zero_()
+            # Padded SWA indices left over from a previous replay would point
+            # into real SWA slots, so set_kv_buffer on padded tokens would
+            # corrupt active requests' KV. Zero the whole buffer so padded
+            # positions map to the sentinel slot (matches piecewise runner).
+            if self.out_cache_loc_swa is not None:
+                self.out_cache_loc_swa.zero_()
             if self.mamba_track_indices is not None:
                 self.mamba_track_indices.zero_()
             if self.mamba_track_mask is not None:

python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py

@@ -238,7 +238,7 @@ class PiecewiseCudaGraphRunner:
                 (self.max_num_tokens,), dtype=self._cache_loc_dtype()
             )
             out_cache_loc_swa = (
-                torch.zeros((self.max_num_tokens,), dtype=torch.int64)
+                torch.zeros((self.max_num_tokens,), dtype=torch.int32)
                 if model_runner.is_hybrid_swa
                 else None
             )